For researchers, especially in remote health monitoring and ubiquitous computing, it is common to expend a significant amount of time and effort to develop data collection systems that can be used outside of the lab or clinic. These systems tend to be customized and highly specific to the task at hand; thus, they are not general enough to support other tasks. In this paper, we present Raproto, an open-source, easy-to-use rapid prototyping platform that facilitates data collection and visualization from sensors on commercially available, off-the-shelf smartwatches. The Raproto platform consists of three components: the smartwatch, communication protocol, and server. These components support the collection, transmission, storage, analysis, and visualization of data. We evaluate our platform on limiting factors including the smartwatch battery life, data loss during transmission, and data latency. Overall, we find that the smartwatch can last for over 24 hours on a single charge, has little to no data loss, and less than a second of data latency per transmission.

Physiological measurements are an integral part of many established and emerging engineering and biomedical applications that involve physiological modeling, physiological state estimation, and physiological closed loop control. In practice, such measurements exhibit a large degree of variability, which is apparent at multiple levels, including disturbances acting on measured signals and unexpected physiological behavior in certain individuals. In this short paper, we present an inference-based approach to estimating the atypicality of an individual's physiological data both at the level of measurement and physiological behavior. For this purpose, we use data from a cohort of subjects to infer, simultaneously, model representations for measurement disturbances and atypicality of physiological behavior. Using a case study on hematocrit (HCT), cardiac output (CO), and mean arterial pressure (MAP) measurements in response to hemorrhage and colloid infusions, we discuss the merits of the presented approach in deriving reliable subject atypicality and signal quality indicators for physiological data.

Recent studies suggest that atypical neural function, due to atypical neural structure, is associated with the behavioral symptoms of Autism Spectrum Disorder (ASD). Additionally, studies suggest that the atypical neural functions and structures associated with ASD change from early childhood to adulthood. This study is the first to develop a multiclass classification model to differentiate neural activity patterns of children and adults with and without ASD depicted by their EEG waveforms. In contrary to the conventional binary classification approaches used in state-of-the-art literature, the multi-class approaches learn the similarity, dissimilarity, common and differentiating patterns among all the categories present in the data. We collected 6 minutes of non-invasive resting-state EEG signals from 105 individuals that include ASD children and adults as well as typical children and adults. Since conventional supervised learning multi-class classifiers suffer from overfitting on limited clinical data, this study employed a few-shot learning mechanism, named prototypical network learning, that is adaptive to limited data and robust against data imbalance issues. Our developed model achieved 85% accuracy in multiclass classification. As the next step, we are developing an interpretable machine learning adaptation for prototypical learning to interpret the model inferences and highlight the brain wave patterns indicative of ASD in different stages of development.

Some medical implants act autonomously: they assess the current health status of a patient and administer treatment when appropriate. An improper treatment, however, can cause serious harm. Here, the decision logic leading to the treatment relies on data obtained from sensors --- an inherently imperfect medium. Coping with these inaccuracies requires the logic to be robust in the sense that slight perturbations in the measurements do not significantly alter the decision. Determining the extent to which an algorithm is robust automatically does not scale well for complex and opaque components. This is particularly problematic when machine learning is involved. Yet, the analysis is feasible for simpler safety-related components such as a runtime monitor, which observes the system and intervenes in a treatment when necessary. Its significantly lower complexity generally allows for providing static guarantees on the runtime behavior of the monitor. Complementing these guarantees with a robustness analysis constitutes a major step toward certifiable medical cyber-physical systems controlled by opaque, machine-learned components. Hence, this paper reports on ongoing research in the direction of a robustness analysis for the runtime monitoring framework RTLola.